|Publication number||US7254647 B2|
|Application number||US 09/816,967|
|Publication date||Aug 7, 2007|
|Filing date||Mar 23, 2001|
|Priority date||Mar 23, 2001|
|Also published as||CN1639679A, EP1370930A2, US20040210687, WO2002077829A2, WO2002077829A3|
|Publication number||09816967, 816967, US 7254647 B2, US 7254647B2, US-B2-7254647, US7254647 B2, US7254647B2|
|Inventors||Gregory J. Mann|
|Original Assignee||International Business Machines Corporation|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (23), Non-Patent Citations (3), Referenced by (7), Classifications (22), Legal Events (5)|
|External Links: USPTO, USPTO Assignment, Espacenet|
1. Field of the Invention
The present invention generally relates to input/output (I/O) data transmission devices, and more particularly to first-in-first-out (FIFO) buffer devices in I/O data transmission paths.
2. Description of the Related Art
InfiniBand (registered Trademark of the InfiniBand Trade Association, Portland, Oreg.) architecture is a new common I/O specification to deliver a channel based, switched-fabric technology that the entire hardware and software industry can adopt. A network and components associated with an InfiniBand network 100 are shown in
Being a fabric-centric, message-based architecture, InfiniBand is ideally suited for clustering, input/output extension, and native attachment in diverse network applications. InfiniBand technology can be used to build remote card cages 15 or connect to attached hosts 35, routers 40, or disk arrays 50. InfiniBand also features enhanced fault isolation, redundancy support, and built-in failover capabilities to provide high network reliability and availability. Featuring high-performance and reliability, these devices provide solutions for a range of network infrastructure components, including servers and storage area networks.
The TCA 10 provides an interface for InfiniBand-type data storage and communication components. Creating InfiniBand adapters that leverage the performance benefits of the InfiniBand architecture is accomplished through a cooperative, coprocessing approach to the design of an InfiniBand and native I/O adapter. The TCA 10 provides a high-performance interface to the InfiniBand fabric, and the host channel communicates with a host based I/O controller using a far less complex interface consisting of queues, shared memory blocks, and doorbells. Together, the TCA and the I/O controller function as an InfiniBand I/O channel deep adapter. The TCA implements the entire mechanism required to move data between queues and to share memory on the host bus and packets on the InfiniBand network in hardware. The combination of hardware-based data movement with optimized queuing and interconnect switch priority arbitration schemes working in parallel with the host based I/O controller functions maximizes the InfiniBand adapter's performance.
The HCA 20 enables connections from a host bus to a dual 1× or 4×InfiniBand network. This allows an existing server to be connected to an InfiniBand network and communicate with other nodes on the InfiniBand fabric. The host bus to InfiniBand HCA integrates a dual InfiniBand interface adapter (physical, link and transport levels), host bus interface, direct memory target access (DMA) engine, and management support. It implements a layered memory structure in which connection-related information is stored in either channel on-device or off-device memory attached directly to the HCA. It features adapter pipeline header and data processing in both directions. Two embedded InfiniBand microprocessors and separate direct memory access (DMA) engines permit concurrent receive and transmit data-path processing.
The interconnect switch 30 can be an 8-port 4×switch that incorporates eight InfiniBand ports and a management interface. Each port can connect to another switch, the TCA 10, or the HCA 20, enabling configuration of multiple servers and peripherals that work together in a high-performance InfiniBand based network. The interconnect switch 30 integrates the physical and link layer for each port and performs filtering, mapping, queuing, and arbitration functions. It includes multicast support, as well as performance and error counters. The management interface connects to a management processor that performs configuration and control functions. The interconnect switch 30 typically can provide a maximum aggregate channel throughput of 64 gigabits, integrates buffer memory, and supports up to four data virtual lanes (VL) and one management VL per port.
To accommodate the different speeds of the data signals being handled, the core logic 210 includes a serialization portion 270 that includes serialization/deserialization (SERDES) units 225. The structure and operation of such serialization/deserialization units is well known to those ordinarily skilled in the art and such will not be discussed in detail herein so as not to unnecessarily obscure the salient features of the invention.
The InfiniBand transmission media 280 is made up of a large number of byte-striped serial transmission lanes 200 that form the links 25. The receive serialization/deserialization units 225 deserialize the signals from the transmission media 280 and perform sufficient conversion to reduce the frequency to one that is acceptable to the core logic 210. For example, if the serialization/deserialization receive units 225 operate to deserialize 10 bits at a time, a 10-to-1 reduction occurs that reduces the 2.5 gigabit per second speed on the transmission media 280 into a 250 MHz frequency that is acceptable to the core logic 210.
The core logic 210 also includes a frequency correction unit 260. The frequency of the signal propagating along the transmission media 280 may not always occur at this wire speed, but instead may be slightly above or below the desired frequency (e.g., by up to 100 parts per million). This inconsistency in the frequency is transferred through the serialization/deserialization units 225. The frequency correction unit 260 includes FIFO buffers 261 that buffer the signal being output by the serialization/deserialization units (SERDES) 225 so as to provide the signal in a uniform 250 MHz frequency to the upper layer logic 250.
The upper link layer logic 250 includes additional FIFO buffers 251 that convert the frequency of the signal output from the frequency correction unit 260 into a frequency that is acceptable to the ASIC 246. During transmission of a signal from the ASIC 246 to the transmission media 280, the process is reversed and the upper link layer logic 250 use FIFO buffers 253. Similarly, the serialization unit 270 uses other transmission serialization/deserialization units 227. Note that no correction is required by the frequency correction unit 262 for signals that are being transmitted to the transmission media 280 because the ASIC 246 generally produces a signal that does not need to be corrected.
One disadvantage of the core logic 210 shown in
In view of the foregoing problems, the present invention has been devised. It is an object of the present invention to provide a parallel-serial architecture network that includes a transmission media and at least one processor connected to the transmission media by a core. The core provides communications between the transmission media and the processor.
The core includes a lower logic layer connected to the processor, serial lanes connecting the lower logic layer to the transmission media, receive and transmit buffers within the serial lanes, and selectors for control of data through the buffers. The receive and transmit buffers correct for fluctuations in the transmission media and alter the frequency of signals being processed along the serial lanes.
The invention may also include serializer/deserializers within the serial lanes. The receive buffers and the transmit buffers are preferably elastic first-in, first-out (FIFO) buffers and the receive buffers and the transmit buffers are both external to the logic layer. The transmit buffers alter a frequency of signals being transferred from the upper layer logic to the transmission media while the receive buffers process signals being transferred from the transmission media to the logic layer. The “processor” can be a host channel adapter, a target channel adapter, or an interconnect switch of the network.
In one embodiment, the invention provides communications between a transmission media and a processor in a parallel-serial architecture. The transmission media operates at a different data speed than the processor. The core includes serial lanes connecting the processor to the transmission media and selectors connected to the serial lanes. The selectors (multiplexors) selectively engage the serial lanes to alter the speed of data passing through the core.
The invention preferably includes a data controller for controlling an operation of the selector. The serial lanes have buffers for performing additional speed alteration of the data. Additional speed adjustments are attained by the selector engaging additional lanes.
The invention includes multiplexors that selectively expand the data lane widths to reduce the clock speed of the data being transferred over the data lanes. By making such a speed adjustment, the invention allows easy communication between a transmission media and a device operating at a different speed.
The foregoing and other objects, aspects and advantages will be better understood from the following detailed description of a preferred embodiment(s) of the invention with reference to the drawings, in which:
As mentioned above, the InfiniBand technologies create a situation where the transmission media is operating at a different speed than some of the devices connected thereto. As shown in
More specifically, elastic buffers 220, 230 reside between the upper link layer logic 250 and the selectors 215, 236 (the receive demultiplexor 215 and the transmit multiplexor 236). The buffers 220, 230 and multiplexors 215, 236 form a low level logic interface 205 frequency correction portion 260 (shown in
To illustrate the utility of the invention, given that 12 data lanes are used in a network, when in a 4×reduction mode of operation, physical data lanes 4-11 can be used as wider extensions of lanes 0-3. Further, when in a 1×mode of operation using these 12 data lanes, FIFOs for data lanes 1-11 can be a wider extension of the FIFO used for data lane 0, thereby achieving up to a 12×speed reduction. Thus, when data is accessed at the transmission media 280 at the InfiniBand data rate, by using the multiplexors 215, 236, the upper link layer 250 can access wider data at a slower rate.
As mentioned above, the invention also produces an advantage in that receive elastic FIFO buffers 220 perform the function of the frequency correction portion 260 and correct any frequency deviations which may occur along the transmission media 280. FIFO buffers 220, 230 also modify the frequency of the signal to that desired by the ASIC 246. Therefore, the FIFO buffers 220 perform the functions that were previously performed by FIFO buffers 251 and 261 shown in
Integration of frequency correction and frequency adjustment processes into the input receive elastic FIFOs 220 also enables the upper layer logic 250 to have clock frequencies that are less than external components connected thereto. For example, the upper layer logic section 250 could have a speed less than 250 MHz while the buffers 220, 230 and serialization 270 portion could operate at approximately 250 MHz (the network shown in
As mentioned above, some hardware in InfiniBand networks have components that operate at different speeds due to different standards imposed. For example, some devices in an InfiniBand network (that operate at 250 MHz) must communicate with non-InfiniBand interface components, such as a component that operates at 62.5 MHz (e.g., PCI-X industry standard, which operates at 133 MHz). These various speed differentials are reconciled the invention. By utilizing the speed reduction capabilities associated with selective lane width (using selectors 215, 236) and integrating the clock-compensation FIFOs 220, the invention improves network performance by lowering the latency of data transmission through the processor.
The upper link layer logic 250 can access the RX and TX FIFOs at a speed down to 62.5 MHz. The speed conversion is accomplished through the lower level logic section 205 that encompasses the selectors 215, 236; the receive and transmit FIFOs 220, 230; and the FIFO data controller 240 that controls lane width using the selectors 215, 236.
The selectors 215, 236 provide the logic to implement lane width changes through the FIFOs 220, 230 respectively in conjunction with the FIFO data controller 240, as discussed in greater detail below in
Logic controller circuitry for pacing the upper transmit layer logic 250 is incorporated therein to prevent FIFO overflow. The logic controller detects when the elastic FIFO buffers 220, 230 are almost full, and then interrupts the clocking of the upper layer logic 250 (pauses data flow) to prevent excessive data flow into these elastic FIFOs 220, 230.
Elastic FIFO buffers 220, 230, each have multiple memory locations into which data is consecutively input. The elastic FIFOs are the preferred form of FIFO used in the invention because they can tolerate different amounts of data (e.g., are expandable). Alternatively, regular FIFOs (e.g., non-elastic) can be used, but with restriction since only a fixed amount of data can be contained within them at any instant in time. Data is output from FIFO's in the same consecutive order in which it is input.
Since these FIFOs 220, 230 are elastic, there are controls on the input that instruct the FIFO buffers to latch the current input and place it into the next memory location, and controls on the output that instruct the FIFO buffers to present the next memory location on the output. There are also indications from the device 220, 230 on how much data is currently in the device. The frequency at which data is removed from the device is not necessarily related to the frequency of data being place into the device, which allows the FIFO to convert the frequency of signals. However, logic controlling the FIFOs must avoid instructing the output to advance to the next entry when there is no data in the device, and avoid instructing the input to place data in the next entry when the device is full of data.
The upper layer clock signal 214 provides a clock from the upper link layer logic 250. This is the step down speed that can operate at a slower speed compared to the media side 280 (e.g., using the multiplexors 215, 236 for the data lane changes by the FIFOs 220, 230). The Data out signal 211 provides a data byte transmitted into the elastic FIFO from below. This is advanced to the next entry in the FIFO when the data_get strobe signal 213 is asserted on an upper layer clock edge. The data_count signal 212 measures the amount of data in FIFO at any given time. This value is increased when data is placed in FIFO from the media side (e.g., multiplexors 215, 236), and decremented when data is removed to the upper layer 250. The data_get strobe signal 213 provides an indication that upper layer logic has fetched the data on data_out 211, and that the FIFO should move to the next entry. The media clock signal 218 is a clock signal operating at the media speed. The data_in signal 216 is a data byte from the media (link) to be placed in FIFO. The data_put signal 217 provides an indication that FIFO should place data on data_in 216 as a signal into FIFO.
The FIFO 220 (230) uses each latching edge of media_clock signal 218 for which data_put_strobe signal 217 is asserted to free an entry in the FIFO, and place the data in the entry on the output of the FIFO. The FIFO uses each latching edge of data_out_clock signal 214 for which the data_byte_get_strobe signal 213 is asserted to place an entry into the FIFO.
The FIFO presents how much data is currently in the FIFO on the data_count 212. This value is updated as data is inserted and removed. The upper layer logic section 250 uses the data_count output 212 to monitor the status of the FIFO. If all of the entries in the FIFO are used, the upper layer logic will deassert data_byte_get_strobe signal 213 until the data_count value indicates that there is an entry available. When the above operation is used, the upper layer logic section 250 can operate at lower frequencies, and clock domain conversion is achieved.
As shown in
Therefore, by monitoring the media lane clock 218, the FIFO data control 240 controls the multiplexors 215, 236 to selectively engage or disengage the transmission lanes 225, 227 so as to perform an appropriate speed reduction to allow the transmission media 280 to properly communicate with the ASIC 246.
For a transmit data path, FIFO data control 240 (or a separate similar unit) can be used to control the removal of data from the transmit elastic FIFOs 230 as shown in
As mentioned above, new technologies in parallel-serial architecture create situations where the parallel-serial architecture transmission media is operating at a different speed than some of the devices connected thereto. In order to accommodate this situation, the invention includes multiplexors 215, 236, connected to elastic FIFOs 220, 230. The multiplexors 215, 236 selectively engage a different number of data lanes 225 (change lane width) in order to perform a speed reduction between the transmission media 280 and the ASIC 246.
Therefore, the invention increases the applicability of parallel-serial architecture transmission media to different speed processing devices by providing different data lane widths to adjust the data speed. As would be known by one ordinarily skilled in the art given this disclosure, the invention is applicable to parallel-serial architecture transmission media that have speeds lower and higher than processors connected to the transmission media. Further, while specific speed reductions are discussed above (e.g., 1×, 4×, etc.), as would be known by one ordinarily skilled in the art given this disclosure, the invention is applicable to any factor of speed adjustment needed, depending upon the specific parallel-serial architecture design involved.
While the invention has been described in terms of preferred embodiments, those skilled in the art will recognize that the invention can be practiced with modification within the spirit and scope of the appended claims.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US4701913||Jun 11, 1986||Oct 20, 1987||Northern Telecom Limited||Circuit and method for extracting signalling information embedded in channelized serial data streams|
|US5175819||Mar 28, 1990||Dec 29, 1992||Integrated Device Technology, Inc.||Cascadable parallel to serial converter using tap shift registers and data shift registers while receiving input data from FIFO buffer|
|US5212686||Sep 29, 1989||May 18, 1993||Plessey Overseas Limited||Asynchronous time division switching arrangement and a method of operating same|
|US5425022 *||Dec 6, 1993||Jun 13, 1995||British Telecommunications Public Limited Company||Data switching nodes|
|US5488408 *||Mar 22, 1994||Jan 30, 1996||A.C. Nielsen Company||Serial data channel metering attachment for metering channels to which a receiver is tuned|
|US5715248||May 21, 1992||Feb 3, 1998||Alcatel Network Systems, Inc.||Derivation of VT group clock from SONET STS-1 payload clock and VT group bus definition|
|US5768546||Jun 5, 1996||Jun 16, 1998||Lg Semicon Co., Ltd.||Method and apparatus for bi-directional transfer of data between two buses with different widths|
|US5938731 *||Jun 23, 1997||Aug 17, 1999||International Business Machines Corporation||Exchanging synchronous data link control (SDLC) frames to adjust speed of data transfer between a client and server|
|US6154797||Aug 10, 1999||Nov 28, 2000||Storage Technology Corporation||System and method for multiplexing serial links|
|US6199137 *||Jan 4, 2000||Mar 6, 2001||Lucent Technolgies, Inc.||Method and device for controlling data flow through an IO controller|
|US6389120 *||Sep 3, 1999||May 14, 2002||Lucent Technologies Inc.||Method and apparatus for multiple logical channel information delivery over multiple suppressed ringing physical channels|
|US6490250 *||Mar 9, 1999||Dec 3, 2002||Conexant Systems, Inc.||Elementary stream multiplexer|
|US6854031 *||Jun 20, 2003||Feb 8, 2005||Cisco Technology, Inc.||Configurable serial interconnection|
|US6956852 *||Jun 25, 1999||Oct 18, 2005||Cisco Technology Inc.||Multi-function high-speed network interface|
|US7054331 *||Sep 13, 2000||May 30, 2006||Intel Corporation||Multi-lane receiver de-skewing|
|US20040213241 *||Dec 29, 2000||Oct 28, 2004||Brana Kukic||Port swapping for inverse multiplexed digital subscriber lines|
|EP0290172A2||Apr 22, 1988||Nov 9, 1988||Advanced Micro Devices, Inc.||Bidirectional fifo with variable byte boundary and data path width change|
|EP0982898A1||Aug 28, 1998||Mar 1, 2000||International Business Machines Corporation||Switching apparatus comprising at least one switch core access element for the attachment of various protocol adapters|
|GB2094523A||Title not available|
|JPH0461528A||Title not available|
|JPS57157348A||Title not available|
|JPS63286033A||Title not available|
|WO1999060495A1||Apr 16, 1999||Nov 25, 1999||Acqiris||Data acquisition system comprising means for analysing and storing in real time|
|1||IBM Technical Disclosure Bulletin, "Data Funnel for Connection of Multiple Channel Types", vol. 32, No. 2, Jul. 1989, pp. 54-55.|
|2||International Search Report, PCT/GB02/01366, Jun. 5, 2003, pp. 1-4.|
|3||Olsen, "Variable-Width FIFO Buffer Sequence Large Data Words", Electronic Design, No. 14, Jun. 1987, pp. 117-120.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US7370132 *||Nov 16, 2005||May 6, 2008||Nvidia Corporation||Logical-to-physical lane assignment to reduce clock power dissipation in a bus having a variable link width|
|US7426597||Sep 16, 2005||Sep 16, 2008||Nvidia Corporation||Apparatus, system, and method for bus link width optimization of a graphics system|
|US7788439||Oct 16, 2008||Aug 31, 2010||Nvidia Corporation||Asymmetrical bus for bus link width optimization of a graphics system|
|US7848394 *||Dec 17, 2004||Dec 7, 2010||Broadcom Corporation||Reconfigurable transceiver architecture for frequency offset generation|
|US8472482 *||Oct 27, 2008||Jun 25, 2013||Cisco Technology, Inc.||Multiple infiniband ports within a higher data rate port using multiplexing|
|US20060056498 *||Dec 17, 2004||Mar 16, 2006||Broadcom Corporation||Reconfigurable transceiver architecture for frequency offset generation|
|US20100103954 *||Oct 27, 2008||Apr 29, 2010||Cisco Technology, Inc.||Multiple Infiniband Ports Within A Higher Data Rate Port Using Multiplexing|
|U.S. Classification||709/250, 709/233, 710/72, 709/234, 710/51, 710/52, 710/71, 710/60|
|International Classification||G06F3/00, G06F13/00, H04L29/10, G06F13/38, H04L13/08, G06F5/00, G06F15/173, G06F13/12, G06F5/06, G06F15/163, H04L12/00, G06F13/40|
|Mar 23, 2001||AS||Assignment|
Owner name: INTERNATIONAL BUSINESS MACHINES CORPORATION, NEW Y
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:MANN, GREGORY J.;REEL/FRAME:011719/0654
Effective date: 20010322
|Jan 29, 2011||FPAY||Fee payment|
Year of fee payment: 4
|Mar 20, 2015||REMI||Maintenance fee reminder mailed|
|Aug 7, 2015||LAPS||Lapse for failure to pay maintenance fees|
|Sep 29, 2015||FP||Expired due to failure to pay maintenance fee|
Effective date: 20150807